Are you annoyed with the anthropomorphizing language being used in the "AI" discourse, but not sure how to talk about this stuff without it?

[mimesatwork@wandering.shop]

@Petesmom @emilymbender You're not wrong, but maybe "extended predictive text" instead of "on steroids" because on steroids refers to how organic creature behaviour changes when it's given extra hormones.

But that's neither here nor there

[jeffcampo@mastodon.social]

@emilymbender I've seen and used "spicy autocomplete" for "LLM"

[jeffcampo@mastodon.social]

@emilymbender also things like "predictive [text] slot-machine" have come up

[martinvermeer@fediscience.org]

@Mimesatwork @Petesmom @emilymbender

'Enhanced'

[martinvermeer@fediscience.org]

@Mimesatwork @Petesmom @emilymbender

It's a necessary, but also difficult uphill battle. After all, a main driver in developing AI is to get computers to do things that, so far, only humans were able to do, either as a challenge or to save on personnel expenditures. Not surprising then if people use the same words to describe what the computer does as what the human would do that they replace.

Like something as basic as 'memory', rather than the more neutral 'storage'. Or in more modern times 'query', i.e., asking questions from the computer, rather than some more neutral term. 'Query' is accepted as describing what you do with an SQL data base, predating AI...

[harib_murshidi@mastodon.social]

@iamnotU @emilymbender We did in many works of scifi though

[martinvermeer@fediscience.org]

@Mimesatwork @Petesmom @emilymbender

Some ideas. For AI in general, we could use 'corpus-based generation', like language, image, action, speech generation. What is characteristic for current AI is how they are built using these large bodies of (often stolen) 'stuff', typically from the Internet.

I have a problem with 'undesirable output' for hallucination. Too vague for me. The undesirability is invariably a failure to be factually correct, which is what users expect from AI, an expectation that is both unreasonable and unwise for a language tool - but also real. Let's make that explicit: a 'factuality violation'.

[frankendrullat@mastodon.social]

@emilymbender @mastodonmigration I have yet to come across a good use case for “AI” for me personally. What they tout as the results of AI has been around since the late 60s. Back then it was called machine learning. Give me a huge data set with a narrow problem and it’s great! Like the examples of finding ancient Maya cities in the jungles of middle America. Yes back then you had to learn LISP to query it ‍️ Even hardware acceleration is not new, remember the Symbolics machines of the early 90s?

[frankendrullat@mastodon.social]

@emilymbender @mastodonmigration The only things that’s new is the natural language interface. And a vastly larger more general dataset to play with, which will / is caus(e/ing) issues as it moves away from the narrow problem and large but specific dataset. Just my 2p.

[tharkonen@mastodon.social]

@emilymbender thank you very, very, very much for this article. So much needed!

[joellmrtz@mastodon.nl]

@emilymbender I love this, thank you so much for sharing. Reminds me of how Anthropic talks about AI models as if they have feelings and as if they can live or die.

[overtondoors@infosec.exchange]

Too generous by half @emilymbender

Stochastic parrot -not- "probabilistic automation" in ley use probabilistic is as opaque as stochastic, but everyone knows parrots don't understand the language they mimic.

Bullshit -in stead of- "hallucination" and "model mistakes" because if it's an response made without the intent to be correct it is by definition bullshit.

[emilymbender@dair-community.social]

@OvertonDoors Cute, that you think I of all people am unaware of the phrase which I coined.

[overtondoors@infosec.exchange]

Well, come on the @emilymbender

It's a good phrase, stick by it.

[emilymbender@dair-community.social]

@OvertonDoors It is a good phrase, but it is a descriptive term for LLMs, not other things called "AI". This is why we proposed "probabilistic automation" for the cases where something more general is needed.

https://medium.com/@emilymenonbender/stochastic-parrots-frequently-unasked-questions-49c2e7d22d11

[overtondoors@infosec.exchange]

eh, you right @emilymbender

Dead right.

The people using these models. They have 1/4 to 1/5 of the functional language skills, reading skills, and ability to reason through the pitfalls inhernt in this tech that you do.

Stochastic Parrot isn't precise. But it's accurate, and in communication for the people who need to hear this accuracy trumps precision.

Anyways, half or most of the well worded criticism -ive read my way through the arguments- boils down to "this phrase is a good counter to our disinformation. Thus I don't like it."

[clickhere@mastodon.ie]

@martinvermeer "Glorified predictive text"?

@Mimesatwork @Petesmom @emilymbender

[martinvermeer@fediscience.org]

@clickhere @Mimesatwork @Petesmom @emilymbender

"Stochastic transmogrifier"

[cobbpg@mastodon.gamedev.place]

@emilymbender I often find it easy to replace “AI” with simply “software” or “automation” depending on the context, because it’s a deliberately vague term to start with. As for chatbots, “text generator” can work just fine. Sometimes I jokingly call them “poetry generators”, because the output has to follow a specific form (e.g. code, slide deck, video script etc.) asked by the user.

[paelnever@masto.es]

@emilymbender
So you say "AI isn't a thing" but keep talking about LLMs as if it was the whole umbrella of technologies under that term.